Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Numba 0.50.1 downloaded from conda fails to compile on GPU (but works when using pip) #5968

Closed
2 tasks done
RemiLehe opened this issue Jul 7, 2020 · 8 comments
Closed
2 tasks done
Labels
bug CUDA CUDA related issue/PR

Comments

@RemiLehe
Copy link

RemiLehe commented Jul 7, 2020

Bug report

When installing numba from conda and running the standard test on GPU, i.e.

conda install numba
python -m numba.runtests numba.cuda.tests

I get a compilation error:

numba.cuda.cudadrv.error.NvvmError: Failed to compile

<unnamed> (25, 17): parse expected binary operation in atomicrmw
NVVM_ERROR_COMPILATION

On the other hand, when using pip:

pip install numba
python -m numba.runtests numba.cuda.tests

the tests are running without any issue.

When playing around (e.g. installing numba always from pip, but installing llvmlite either from conda or pip), it seemed to me that the issue comes from the version of llvmlite downloaded from conda...

More details

In the case of conda, the following packages were downloaded.

  llvmlite           pkgs/main/linux-64::llvmlite-0.33.0-py37hd408876_0
  numba              pkgs/main/linux-64::numba-0.50.1-py37h0573a6f_0
  tbb                pkgs/main/linux-64::tbb-2020.0-hfd86e86_0

On the other hand, in the case of pip, the following packages were downloaded:

llvmlite-0.33.0-cp37-cp37m-manylinux1_x86_64.whl
numba-0.50.1-cp37-cp37m-manylinux2014_x86_64.whl

The tests are run on a GTX 1080 Ti with cudatoolkit 9.0 installed from conda (Build h13b8566_0).

@stuartarchibald
Copy link
Contributor

Thanks for the report, I can reproduce. From the error I think the Anaconda llvmdev builds aren't carrying the patch that prevents instruction upgrade.

@RemiLehe
Copy link
Author

RemiLehe commented Jul 8, 2020

Thanks for looking into this.
Is there a chance that the conda build could be updated, so as to include the patch?

@seibert
Copy link
Contributor

seibert commented Jul 8, 2020

We're checking with them now to get a fixed build rolled out.

@stuartarchibald
Copy link
Contributor

CC @angloyna for updates (with thanks for patching upstream!)

@angloyna
Copy link

Anaconda llvmdev has been patched, the latest numba 0.50.1 should have it. Let me know if we're still seeing the issue.

@RemiLehe
Copy link
Author

Great! I confirm that this fixed the issue. Thanks a lot!

@stuartarchibald
Copy link
Contributor

Thanks to you both @angloyna @RemiLehe, good to see this fixed.

@RemiLehe
Copy link
Author

As a quick follow-up: I just noticed that, even though llvmlite 0.33 was updated in the numba channel, it is not updated in the conda-forge channel.
Therefore, users that were using e.g. conda install -c conda-forge numba were still getting the above-mentioned bug.

Is there any chance that the conda-forge package could be updated too?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
bug CUDA CUDA related issue/PR
Projects
None yet
Development

No branches or pull requests

4 participants